Google Willow tests quantum contextuality up to 105 qubits | Magica
Google’s Willow experiments turn quantum contextuality into a hardware test
Editorial Team
••📖7 min read
Google Quantum AI reports tests of quantum contextuality spanning up to 105 qubits on Willow processors. The experiments expose hardware errors and beat restricted classical strategies, while the largest task relies on an unproven classical benchmark.
A four-qubit magic-square game won 98.3% of trials, above the 88.9% ceiling for classical players who cannot communicate.
Separate experiments scaled to 71 and 105 qubits; the largest task’s classical comparison remains conjectural.
Google says the structured tests exposed systematic hardware errors that random-circuit benchmarks can miss.
Google Quantum AI, Google’s quantum-computing research group, reports experiments spanning up to 105 qubits that turn quantum contextuality into a test of processor performance. Announced November 19 alongside publication in Science, the work uses Willow superconducting processors to test whether quantum correlations survive as more qubits—the units of quantum computation—work together.
The practical finding is a hardware diagnostic: Google says these structured tasks exposed subtle errors that its conventional random-circuit checks could miss. The results also beat classical strategies under specific restrictions, such as forbidding communication between players or limiting circuit depth. They do not demonstrate a useful computation running faster or more cheaply than on an ordinary computer. Google’s announcement points to chemistry and materials science as potential beneficiaries of better hardware.
This is a new publication milestone for an existing research effort. The study first appeared as a preprint on December 1, 2025. Willow itself was introduced in December 2024 with error-correction and random-circuit-sampling results. Google Quantum AI, established in 2012 to build a useful large-scale quantum computer, is testing hardware it develops: Google says it fabricated Willow at its Santa Barbara facility. Its Willow introduction explicitly distinguished computational benchmarks from commercially useful applications.
Google’s company-reported four-qubit magic-square results averaged 98.3% wins, compared with the 88.89% classical ceiling for players who cannot communicate. Credit: Google Quantum AI. Source: Google Research.
Epic says it is using AI to rewrite parts of Unreal Engine and plans to let developers connect their preferred models to UE6. The strategy combines claimed maintenance gains with a broader overhaul of programming, distribution and portable game assets, while rival Unity already offers outside-model connections.
Editorial Team
What the four-qubit win rate establishes
Contextuality means that quantum measurement outcomes cannot all be explained by values fixed in advance independently of which other compatible measurements accompany them. Theory identifies it as a resource for certain quantum algorithms. Measuring it on a processor helps examine that resource without establishing a general computational advantage.
The most accessible demonstration is the Mermin–Peres magic-square game. Two players receive a row and a column of a three-by-three grid. Each fills their assigned cells with +1 or −1. Every row must multiply to +1 and every column to −1; the players win if their entries agree at the intersection. They may coordinate beforehand but cannot communicate after receiving their assignments.
A single fixed grid cannot satisfy those rules: the rows require an even total number of minus signs, while the columns require an odd total. For nine equally likely row-column assignments, classical players can therefore win at most eight, or 88.89%. Sharing two entangled qubit pairs allows quantum players, ideally, to win every time.
Google’s experimental account reports 100,000 trials for each assignment, or 900,000 in total. The average winning probability was 0.9830 ± 0.0001, about 9.4 percentage points above the classical ceiling. This used four qubits, two per player. The 98.3% figure is not the result of a 105-qubit game.
The measurement procedure matters. In the version producing that result, each player measured two answers and inferred the third from the multiplication rule. In a separate comparison in the preprint’s methods, measuring all three directly required more gates and produced more errors: the win rate was 93.55%, versus 97.73% for the inference-based version in that comparison, with 10,000 trials per assignment for each. Both remained above eight ninths, but their different demands show why the score must be tied to the protocol.
A second small-system test used sequential measurements on two qubits and a third helper qubit to record outcomes. It tested a Bell–Kochen–Specker inequality: models assigning context-independent values are bounded by a score of 4, while quantum mechanics permits 6. Google’s November account reports 5.590 ± 0.008; the December 2025 preprint reports 5.618 ± 0.005. The two accounts do not explain that difference, and both exceed the same bound. The inequality is independent of the initially prepared quantum state, provided its terms are evaluated for the same preparation.
Company-reported GHZ-game results stay above the classical winning probability through 71 players, with a narrowing margin. This differs from the separate certification of genuine entanglement across 45 qubits. Credit: Google Quantum AI. Source: Google Research.
Larger systems face different limits
The methods identify both 72-qubit and 105-qubit Willow processors. The larger coordination game reached 71 players, with one qubit per player. It used a GHZ state, a shared quantum superposition of all qubits being zero and all being one. Each player received a binary question and returned a binary answer without communicating; the group won when the answers’ combined parity matched the rule set by the questions.
The classical ceiling changes with the number of players: it is 75% for three and approaches 50% as the group grows. Google reports staying above that ceiling through 71 players, although the margin narrowed with noise. The preprint describes 20 question sets and 50 randomized state-growth patterns at each tested size, with 1,000 repetitions per configuration.
Google separately reports verifying genuine entanglement across 45 qubits. As an account of the experiments explains, that certification and the 71-player win rate answer different questions. Beating a game’s classical strategy does not establish that every participating qubit has been certified as part of one genuinely entangled state.
The 105-qubit experiment tackled a different task, the two-dimensional hidden linear function problem. An input matrix specifies mathematical relationships, and the processor returns a binary string satisfying them. A 2018 theoretical result established that sufficiently large instances separate shallow quantum circuits from restricted classical circuits. “Shallow” refers to a fixed number of operation layers as the problem grows, rather than a fixed elapsed runtime.
Google reports valid answers more than half the time at the largest size. The preprint describes 1,000 random problem instances at each tested size, from 9 to 105 qubits, measured 100 times apiece. Outputs were checked with classical simulations of the quantum circuits. The experiment thus compares constrained computational resources; it does not show that classical computers cannot solve these instances.
The quantum implementation used four layers of entangling gates. To account for errors, the researchers divided four by the measured success probability, reporting an effective depth of about six at 105 qubits. This is a resource metric incorporating repeated attempts, rather than a measurement of elapsed computing time or deployment cost.
The decisive caveat is on the classical side. The authors say the established asymptotic proof does not supply a usable bound for these small, noisy experiments. They instead propose a conjectured comparison counting layers of classical two-input, one-output gates, without restricting connectivity. The methods note a further mismatch: the earlier theorem requires success above seven eighths, while the measured success rate falls below that level at about 35 qubits. A finite-size classical bound that matches the experimental success criterion remains necessary to establish the separation claimed for this task.
Google’s company-reported hidden-function results reach about 0.6 success probability at 105 qubits. The gray classical comparison is conjectured; this plot does not establish general computational advantage. Credit: Google Quantum AI. Source: Google Research.
Why the hardware test may matter sooner
Contextuality experiments have precedents beyond Willow. A 2022 magic-square experiment used photon pairs entangled in polarization and orbital angular momentum, under locality and fair-sampling assumptions. Google’s contribution is to extend related tasks across a programmable superconducting platform and use their sensitivity to probe hardware performance.
The company says the larger structured tests exposed coherent errors—systematic deviations in quantum operations—that random circuits tend to average away. The methods illustrate the calibration demands with the smaller inequality test: measuring and resetting the helper qubit required frequency excursions that accumulated unwanted phase, which the researchers compensated with calibrated rotations.
For the games and inequality, the ideal quantum target is fixed: a win probability of one or an inequality score of six. As an analysis of the results notes, distance from that ideal can measure hardware progress even as classical algorithms improve. That fixed target should not be confused with the conjectured classical bound for the hidden-function task.
The next test is whether the diagnostic improves computation
The open mathematical issue is a proven classical comparison for the hidden-function experiment at the tested sizes and error levels. The engineering question is whether the errors these benchmarks expose can be reduced in ways that improve larger structured computations.
Google’s released data package includes game, inequality, entanglement-fidelity and hidden-function datasets, plus a notebook for reproducing the principal figures. It gives researchers material to examine the measurements and comparison choices.
Commercial usefulness requires further evidence. Google presents chemistry and materials science as prospective uses of improved hardware; a report on the study notes that these experiments simulated neither. The next consequential demonstration would connect the diagnostic to better performance on an application, with a classical comparison and resource accounting appropriate to that application.
At Ignite, Nvidia and Microsoft are expanding local coding AI, cloud desktop streaming and Azure infrastructure. The announcements build on an existing Windows AI strategy, with AMD alternatives, staggered availability and no comparable deployment-cost results in the announcements.
Microsoft and Nvidia are expanding Azure’s Vera Rubin virtual-machine preview, with systems already running in its data centers. Microsoft’s separate confidential-AI showcase covers four GPUs, while the announcements leave pricing, regions and general availability unspecified.
Meta is reconsidering its 2030 climate deadline after net emissions reached 11.8 million tonnes in 2025. Hardware and infrastructure purchases dominate its reported footprint, while new gas generation adds a longer-term challenge.
Three House committee chairs want the Remote Access Security Act in the annual defense bill as scrutiny of Nvidia intensifies. The available Senate text would authorize controls on specified foreign users’ remote access, with computing limits left to Commerce.
A bipartisan House disclosure shows requests to sell 3.4 million AI chips to China and approvals for more than 879,000. The oversight question is whether Commerce verified the buyers, supply protections and access restrictions behind those authorizations.
OpenAI says its internal Performance Factory tracked more than 1,700 investigations in just over three months. About 400 cases reached proposed fixes, but far fewer shipped repairs used Codex’s code unchanged, and engineers retain approval.
Activists protested Anthropic’s stock-market debut outside Nasdaq, with at least one arrest and conflicting accounts of the total. Their campaign targets retirement-fund exposure to AI shares, alongside concerns about jobs, data centers and public accountability.
Anthropic says cash, growing revenue and multiyear payment schedules support its vast compute commitments. Its filing shows why flexibility matters: short-term capacity costs more, while major cloud contracts require payment regardless of usage.
OpenAI is rolling out live transcription for dictated ChatGPT prompts across free and paid plans. Users can see words as they speak and review the text before sending, but the company has not supplied performance measurements or a replacement recording maximum.
Boltz has raised a Bessemer-led $110 million Series A to expand molecular-design tools sold by usage. Its open-source research brought broad adoption; the commercial test is whether paid computation delivers repeatable gains in laboratory discovery.
Donald Trump said he and Xi Jinping supported minimal AI guardrails but had not discussed definitions. Their Shenzhen meeting leaves unresolved how bilateral safety cooperation would coexist with U.S. opposition to global regulation.
Gemini can now click and type across Mac apps and websites in a background window, Google says. The paid rollout expands its desktop automation, but app permissions expose active accounts and file changes that backups cannot always undo.
The House AI Safety and Innovation Caucus promises legislative advocacy and rapid incident briefings. Its members already have bipartisan safety proposals in play, making the choice of safeguards its next test.
Binance Wallet says it will return all subscription commitments, distribute roughly 2,957 Anthropic tokenized shares free and add a $1 million token giveaway after demand far exceeded its allocation.
Early Anthropic investor Dustin Moskovitz is urging frontier AI developers to slow down. He and Cari Tuna moved their stake to a nonprofit in early 2025, directing potential returns toward philanthropy while leaving questions about the holding's value and governance rights.
A national study finds higher odds of hosting data centers in majority-Black, Hispanic and Asian neighborhoods. Facility types, zoning rules and unequal access to public hearings complicate who bears the costs and receives the benefits.
A debut-day estimate values the Anthropic shares FTX sold in 2024 at $49.7 billion. The comparison exposes the scale of subsequent paper gains, but contrasts cash raised for bankruptcy repayments with a hypothetical stock-market valuation.
Anthropic raised $81.3 billion in a record IPO and opened 34% above its offer price before retreating. Rapid revenue growth and adjusted operating profits sit alongside vast computing commitments and limited investor control.
The Kimi developer reportedly abandoned planned Nvidia B300 capacity for 2027 after new US restrictions. The deal exposes financing risks for overseas AI clouds, but neither a model delay nor a $150 million cash loss is established.